1 files changed, 51 insertions, 0 deletions

diff --git a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
index 0442adc83..a27f2f73a 100644
--- a/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
+++ b/src/shader_recompiler/backend/spirv/spirv_emit_context.cpp
@@ -480,6 +480,7 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
     DefineTextures(program.info, texture_binding, bindings.texture_scaling_index);
     DefineImages(program.info, image_binding, bindings.image_scaling_index);
     DefineAttributeMemAccess(program.info);
+    DefineWriteStorageCasLoopFunction(program.info);
     DefineGlobalMemoryFunctions(program.info);
     DefineRescalingInput(program.info);
     DefineRenderArea(program.info);
@@ -877,6 +878,56 @@ void EmitContext::DefineAttributeMemAccess(const Info& info) {
     }
 }
 
+void EmitContext::DefineWriteStorageCasLoopFunction(const Info& info) {
+    if (profile.support_int8 && profile.support_int16) {
+        return;
+    }
+    if (!info.uses_int8 && !info.uses_int16) {
+        return;
+    }
+
+    AddCapability(spv::Capability::VariablePointersStorageBuffer);
+
+    const Id ptr_type{TypePointer(spv::StorageClass::StorageBuffer, U32[1])};
+    const Id func_type{TypeFunction(void_id, ptr_type, U32[1], U32[1], U32[1])};
+    const Id func{OpFunction(void_id, spv::FunctionControlMask::MaskNone, func_type)};
+    const Id pointer{OpFunctionParameter(ptr_type)};
+    const Id value{OpFunctionParameter(U32[1])};
+    const Id bit_offset{OpFunctionParameter(U32[1])};
+    const Id bit_count{OpFunctionParameter(U32[1])};
+
+    AddLabel();
+    const Id scope_device{Const(1u)};
+    const Id ordering_relaxed{u32_zero_value};
+    const Id body_label{OpLabel()};
+    const Id continue_label{OpLabel()};
+    const Id endloop_label{OpLabel()};
+    const Id beginloop_label{OpLabel()};
+    OpBranch(beginloop_label);
+
+    AddLabel(beginloop_label);
+    OpLoopMerge(endloop_label, continue_label, spv::LoopControlMask::MaskNone);
+    OpBranch(body_label);
+
+    AddLabel(body_label);
+    const Id expected_value{OpLoad(U32[1], pointer)};
+    const Id desired_value{OpBitFieldInsert(U32[1], expected_value, value, bit_offset, bit_count)};
+    const Id actual_value{OpAtomicCompareExchange(U32[1], pointer, scope_device, ordering_relaxed,
+                                                  ordering_relaxed, desired_value, expected_value)};
+    const Id store_successful{OpIEqual(U1, expected_value, actual_value)};
+    OpBranchConditional(store_successful, endloop_label, continue_label);
+
+    AddLabel(endloop_label);
+    OpReturn();
+
+    AddLabel(continue_label);
+    OpBranch(beginloop_label);
+
+    OpFunctionEnd();
+
+    write_storage_cas_loop_func = func;
+}
+
 void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
     if (!info.uses_global_memory || !profile.support_int64) {
         return;